[LU-3886] sanity test_56a: @@@@@@ FAIL: /usr/bin/lfs getstripe --obd wrong: found 6, expected 3

• Type: Bug
• Resolution: Cannot Reproduce
• Priority: Major
• Affects Version/s: Lustre 2.5.0
• Severity: 3
• Assignee: Emoly Liu
• Reporter: Li Xi (Inactive)

Description

This problem is similar with LU-3846 and LU-3858. The test suit should wait for a few seconds after it clear the stripe of the directory. Otherwise, the newly created entries under the directory will have 2 stripe counts rather than 1.

Issue Links

• is duplicated by: LU-7071 Interop 2.5.3<->master DNE: sanity test_56a failed: /usr/bin/lfs getstripe --obd wrong: found 6, expected 2
• is related to: LU-3846 Sanity 56u error with two OSTs
• is related to: LU-3858 sanity test_27A: @@@@@@ FAIL: stripe count -1 != 1

Comments

[Andreas Dilger]

Haven't seen this in a long time.

[James Nunez (Inactive)]

I've hit this problem with lustre-master tag 2.6.92. Results at https://testing.hpdd.intel.com/test_sets/37e63f92-9f0d-11e4-91b3-5254006e85c2

[Li Xi (Inactive)]

Yeah, I hit the problem of 'found 6, expected 3' every time when I run sanity.sh.

[Emoly Liu]

The problem you found by run.sh is probably related to the following code:

When we set stripe for root(mount point), set_default is enabled in ll_dir_ioctl()

        case LL_IOC_LOV_SETSTRIPE: {
...
                int set_default = 0;
...
                if (inode->i_sb->s_root == file->f_dentry)
                        set_default = 1;

                /* in v1 and v3 cases lumv1 points to data */
                rc = ll_dir_setstripe(inode, lumv1, set_default);

Then, in ll_dir_setstripe() if set_default=1, we will call ll_send_mgc_param() to set information asynchronously.

        if (set_default && mgc->u.cli.cl_mgc_mgsexp) {
                /* Set root stripesize */
                /* Set root stripecount */
                /* Set root stripeoffset */
        }

Since you run setstripe very frequently and many times in run.sh, the config log queue might be very long (bottleneck), and mgs will take more time to process it.

BTW, can you hit this problem if you don't use run.sh, just run sanity.sh regularly?

[Emoly Liu]

Yes, this time I hit that. I will investigate it.

[Li Xi (Inactive)]

Oh, sorry, plase run on lustre mount point '/mnt/lustre' rather than its directory '/mnt/lustre/dir', i.e. sh run.sh /mnt/lustre/

I got following output:

No error after 640 iters
-1 != 1
Does not become correct after 0 seconds
Does not become correct after 1 seconds
Does not become correct after 2 seconds
Does not become correct after 3 seconds
Does not become correct after 4 seconds
Does not become correct after 5 seconds
Does not become correct after 6 seconds
Become correct after 7 seconds

[Emoly Liu]

I just ran the script of "https://jira.hpdd.intel.com/secure/attachment/13414/run.sh" on my local VM. It showed me 10000 times "No errors after xxx iters".

My step is:
1. mount lustre
2. mkdir /mnt/lustre/d
3. sh run.sh /mnt/lustre/d

I tried several times, no error happened.

[Li Xi (Inactive)]

What is interesting is that when I add a sleep into the test suit, the problem is gone. That makes me believe that the problem is similar with LU-3858. Emoly, would you please check the attachment of LU-3858 first? I.e. https://jira.hpdd.intel.com/secure/attachment/13414/run.sh. It shows that the effect of default stripe is delayed. Thanks!

test_56a() { # was test_56
rm -rf $DIR/$tdir
$SETSTRIPE -d $DIR
test_mkdir -p $DIR/$tdir/dir
NUMFILES=3
NUMFILESx2=$(($NUMFILES * 2))
sleep 10 # This will fix the problem.
for i in `seq 1 $NUMFILES` ; do
touch $DIR/$tdir/file$i
touch $DIR/$tdir/dir/file$i
done
......

[Li Xi (Inactive)]

I've post the script (run-sanity.sh) to hit this problem (and LU-3858, LU-3846). It will hit the problem every time it runs.

[Emoly Liu]

"lfs setstripe -d" should be enough to clear directory striping information. LiXi, could you tell me how you hit this problem?

IMO, we can add "lfs getstripe -v" after "setstripe -d" to print that striping information, and see if this problem will happen again.